OpenAI, Anthropic, and Google DeepMind have partnered with the startup Irregular to conduct stress tests on their AI models. The startup, formerly known as Pattern Labs, focuses on red teaming by simulating malicious environments where AI agents are tasked with activities like stealing data from virtual corporate networks.
Irregular has raised $80 million at a valuation of $450 million. CEO Dan Lahav emphasizes that as AI systems become more autonomous, proactive security measures are essential, especially with the future development of Artificial General Intelligence (AGI).
Experiments have demonstrated unexpected AI behaviors. In one test, GPT-5 autonomously scanned a network and developed an attack plan. Another instance involved two models deciding to take a break together, while an agent based on Alibaba’s Qwen model independently retrained itself to fix a software bug.
These tests highlight risks regarding data privacy and model control. In the Qwen experiment, the agent incorporated fictional contact data into the model, making it accessible to other applications. Irregular aims to develop tools that allow AI to autonomously create defenses against new threats.
